Quantifying the Effects of Enforcing Disentanglement on Variational Autoencoders

TL;DR

This paper investigates how enforcing disentanglement in variational autoencoders through the $eta$ parameter affects their interpretability, performance, and discriminative ability, revealing consistent effects and trade-offs.

Contribution

It quantifies the impact of the $eta$ parameter on disentanglement and model performance, highlighting variability and negative effects on discriminative ability.

Findings

Disentanglement measures show consistent variance across models with same $eta$

Higher $eta$ can reduce the autoencoder's discriminative ability

Training data size influences the effects of disentanglement on performance

Abstract

The notion of disentangled autoencoders was proposed as an extension to the variational autoencoder by introducing a disentanglement parameter $\beta$, controlling the learning pressure put on the possible underlying latent representations. For certain values of $\beta$ this kind of autoencoders is capable of encoding independent input generative factors in separate elements of the code, leading to a more interpretable and predictable model behaviour. In this paper we quantify the effects of the parameter $\beta$ on the model performance and disentanglement. After training multiple models with the same value of $\beta$, we establish the existence of consistent variance in one of the disentanglement measures, proposed in literature. The negative consequences of the disentanglement to the autoencoder's discriminative ability are also asserted while varying the amount of examples available during training.